D‐S Evidence Theory Integrated Filters Guided Adaptive Molecular Force Search Optimizer for Feature Selection of Microarray Cancer Datasets

Cancer gene expression data possess the characteristics of high dimensionality, a small sample size, and class imbalance, which pose significant challenges for the precise diagnosis of cancer subtypes. This paper proposes the Dempster–Shafer theory (D‐S) evidence theory integrated filters guided adaptive molecular force search optimizer (DS‐EFG‐AMFSO), which is used for feature selection (FS) in microarray cancer data. The core innovation of this method lies in the utilization of the D‐S evidence theory to integrate four different FS methods, thereby dividing the evaluation results into two levels—an initial screening subset that retains approximately 89% of the peak features and an elite subset that contains approximately 5% of the peak features. The former is used to narrow the search space, while the latter guides the subsequent optimization direction. Second, in three stages of molecular force search, three complementary optimization strategies are introduced, respectively. In the exploration stage, a Gaussian penalty factor is used to enhance the global search ability and avoid getting stuck in local optima; in the transition stage, a triangular tracking strategy is designed to balance the exploration and exploitation processes; in the exploitation stage, the elite subset is used to guide the search to converge towards high‐quality feature subsets, improving the local optimization accuracy. To verify the effectiveness of the method, systematic experiments were conducted on 16 microarray cancer gene expression datasets. First, by comparing with the basic method, the contributions of three optimization strategies to the algorithm performance were confirmed. Subsequently, DS‐EFG‐AMFSO was compared with nine advanced intelligent FS methods. The experimental results show that DS‐EFG‐AMFSO significantly outperforms the comparison methods in multiple indicators such as fitness value, classification accuracy, precision, recall rate, F‐score, and the number of selected features. Especially on 11 datasets, this method achieved the best results in terms of precision, recall rate, and F‐score and demonstrated a significant advantage in the average number of selected features, indicating that it can effectively reduce the feature dimension while maintaining high classification performance.